OpenAI Launches dots: Always-On AI Agents That Work While You Don't

On September 29, OpenAI unveiled dots at DevDay 2026. The official line: “remarkably capable, always-on agents built to handle everything.”

dots is a colleague that lives in the cloud. Give it a goal and it works on its own; close the tab and it keeps going; come back and check the results.

One dot = One Cloud Computer + One Browser

Each dot runs on GPT-6 Astra. OpenAI gives every dot a dedicated cloud computer and a browser. The dot works on that machine, and you can open it anytime to check progress.

The workflow is three steps: set a goal, connect apps, set rules.

You tell the dot what to achieve — say, “watch this app’s user feedback and fix bugs.” The dot learns your preferences from feedback and gets smoother with use. OpenAI’s example: an early tester’s dot noticed an invoice never sent to a partner, prepared it itself, and only sent it after the tester nodded.

dots also connects apps. Through ChatGPT’s plugin ecosystem it reaches 4,000+ apps. Talk to it or call it from ChatGPT’s web, desktop, and mobile apps — or message it in Slack and Teams. SMS support is still being built.

Who Gets It First, and What It Costs

dots goes to paying users first: ChatGPT Pro and Business Premium subscribers, rolling out by market. Pro isn’t available in the EEA, Switzerland, or the UK for now. Enterprise, education, and healthcare workspaces can trial the beta, but it’s off by default — admins must enable it manually.

Pricing: the first dot is included in Pro or Business Premium — no extra charge. For the first month after launch, usage caps get looser too. Chatting with a dot doesn’t count against ChatGPT quotas; tasks it starts in Codex or ChatGPT Work count as usual.

Org users also get “expert dots,” each with its own identity, credentials, and system permissions. Piloting in enterprises first; OpenAI is working with Microsoft to plug expert dots into Agent 365’s governance and security stack.

Safety: Custom Rules + Read-Only Background

The biggest question about agents is whether they’ll go rogue. OpenAI offers several layers of restraint.

First, Custom Rules. You define what a dot may do on its own, what must ask you first, and what it never does. Password changes, for example, always stay with the user.

Second, automatic review. Actions that could affect accounts or involve information sharing get checked first against your instructions, rules, and OpenAI’s safety requirements.

Third, background permissions. When you’re not interacting, dots do “proactive research” — read-only in that mode. No sending messages, no modifying app content, no controlling your browser or computer.

Fourth, transparency. Every dot’s cloud computer is open for inspection anytime. Connecting your own computer is optional and off by default. Passwords saved on websites aren’t exposed to the model.

OpenAI keeps a backstop: its monitoring systems can pause or shut down a dot on safety signals. The official line: dots will still make mistakes.

Rivals Got There First

On September 8, Meta launched the personal agent Muse in the US. Google moved earlier — the May I/O conference already showed Gemini Spark, another 24/7 personal agent.

Different plays. Muse leads with the consumer market, riding Facebook, Instagram, and WhatsApp traffic. dots clearly leans toward work: the plugin ecosystem, Slack and Teams integration, the Microsoft enterprise partnership — all aimed at the office.

Also at DevDay: Pro 500 and GPT-6.1 Sol

DevDay brought more:

• Pro 500: a new $500/month tier with 25x Plus usage.
• Ultrafast: a new speed tier — 300 tokens/sec in Codex, 8x standard speed.
• GPT-6.1 Sol: the upgraded GPT-6 Sol, available to Plus through Edu users in ChatGPT Work and Codex. API pricing: $2 input / $10 output per million tokens.
• ChatGPT Space: team collaboration spaces where teammates, ChatGPT, and your dot work together.
• Pages: documents co-edited by humans and agents.

OpenAI also shared a number: ChatGPT now has 1.2 billion weekly users.

The Takeaway

ChatGPT solves “you ask, it answers.” dots aims at “you set the goal, it delivers the result.” OpenAI already uses it internally: a bug report appears in Slack, the dot investigates on its own, fixes it, opens a PR — with a video of the changes attached.

The real test is permissions and trust. The more capable the agent, the harder users must think about what can be fully delegated and what needs human sign-off. Custom Rules and automatic review are OpenAI’s answer; whether they suffice depends on what the first users report back.

Sources: OpenAI’s official announcement, verified across Betanews, MartechAI, Seoul Economic Daily, and others.